In mechanical signal processing, most mechanical vibration signals are modulation signals and the observation signals detected by sensors are always the aggregation of independent vibration sources. The mechanical fault diagnosis method based on singular values of blind separated sources' envelope matrixes was proposed. This proposed method was composed of the blind source separation of multi-channel sensor signals, the envelope analysis of independent vibration sources, the singular value decomposition of envelope matrixes, the feature extraction of singular values and the pattern recognition with the LS-SVM classifiers. The experimental results of hydraulic gear pump indicate that this proposed method is effective. The feature vectors of singular values, with favorable numerical stability, possess well characteristics of clustering and partitioning. The recognition ratio of the LS-SVM classifiers is much higher. Then this method can be applied to signal processing and fault diagnosis of mechanical equipment.